Mendral cuts costs after Opus upgrade

The real lesson is that frontier models can reduce spend when they become the planner, not the workhorse. Better orchestration beats brute-force prompting.

• –Mendral reports that 4 out of 5 failures never reach Opus, so most volume is handled by a cheap classifier instead of the premium model
• –The system uses Opus for hypothesis generation and Haiku for log reading and targeted retrieval, which keeps the expensive context clean
• –Pushing logs into prompts is replaced by SQL against ClickHouse, which is a more scalable pattern for high-volume debugging data
• –Capping sub-agent fan-out matters as much as model choice, because runaway delegation can erase any token savings
• –The architecture is most relevant for teams handling noisy event streams like CI, security alerts, or telemetry, not just software build failures